pixi and micromamba install different builds of the same package

[Yura52]

Hi! Namely, this:

pixi init a
cd a
pixi add pytorch

results in pytorch=2.4.0=cpu_mkl_py312h31352b0_101. While this:

micromamba env create -n b python=3.12.6 pytorch

results in pytorch=2.4.0=cuda120_py312h26b3cf7_301. Is this behavior expected? Can I somehow tell Pixi to follow conda/mamba/micromamba when choosing builds?

I tested the above on linux-64 with pixi=0.31.0 and micromamba=1.5.9.

P.S. Also, I am curious which of the two builds is actually better/faster?

[baszalmstra]

Did you add the cuda system requirement?

[ruben-arts]

This does fix it:

[project]
channels = ["conda-forge"]
name = "a"
platforms = ["linux-64"]

[system-requirements]
cuda = "12.1" # <<<

[dependencies]
pytorch = ">=2.4.0,<3"

> pixi list --no-install
....
pytorch               2.4.0       cuda120_py312h26b3cf7_301  24.6 MiB   conda  pytorch-2.4.0-cuda120_py312h26b3cf7_301.conda
...

[Yura52]

Thanks!

[baszalmstra]

P.S. Also, I am curious which of the two builds is actually better/faster?

If you have a CUDA capable GPU then the cuda version is usually faster.

[Yura52]

I forgot to clarify that I implied the CPU-related speed. Initially this small investigation was motivated by this issue: I was wondering whether the PyTorch build that I am using with Pixi is the fastest ones for CPU-based computations.

[baszalmstra]

I dont think there are micro-arch optimized builds for pytorch so the one you mentioned is probably the fastest for any CPU.